Multimodal evaluation of small bowel crohn’s disease activity integrating clinical data, CTE radiomics, and deep learning on DBE

This study aimed to develop a multimodal model for assessing small bowel Crohn’s disease (SB-CD) activity by integrating clinical data, radiomic features from computed tomography enterography (CTE), and deep learning analysis of double-balloon enteroscopy (DBE) images. Uni‑/multivariate logistic regression identified clinical predictors. Radiomic features extracted from semi-automatically segmented CTE images were selected using LASSO regression. DBE images were processed via ResNet‑18 for classification and deep‑learning feature extraction. The clinical‑radiomic‑deep learning multimodal models were developed. Performance was evaluated with receiver operating characteristic (ROC), calibration, decision curve analysis (DCA), nomograms, and Shapley Additive Explanations (SHAP) plots; DeLong, net reclassification improvement (NRI) and integrated discrimination improvement (IDI) tests compared models. Four clinical factors, seven radiomic features, the binary deep learning classification output, and six deep learning-derived features were selected to construct the models. Overall, single-modality models demonstrated lower performance compared to multimodal models, with all seven models achieving area under the ROC curve (AUC) values exceeding 0.90 in the training cohort. Multimodal models (Combined 2 & 3) yielded the highest performance, outperforming single‑modality and clinical‑radiomic models. Nomograms visualized individualized risk, and SHAP plots highlighted key features. The superiority of multimodal integration was confirmed by Delong, NRI, and IDI analyses. The multimodal model integrating clinical data, radiomic features, and deep learning outputs showed potential for assessing disease activity in SB-CD and may provide complementary information for clinical evaluation.